> 5) Forget it (this proposal is the easiest to write up :-).
This could be very harmful for the MPIO interface. Notice currently
there are far more sequential machines than parallel machines and
zillion times more software using sequential I/O than MPIO.
Scientists using MPI to calculate a complex problem would like to
write the result to a file for visualization and analysis by other
existing packages. Similarly, they also may want their newly coded
MPI application to read data files from other sequential
application. Without a defined canonical format, they are stuck.
Their only alternative would be to funnel all data through one
process which does I/O in the traditional way. If the MPIO interface
cannot interact with sequential I/O application, it will be living on
an island.
Someone has suggested to use the native machine format for data exchange
since majority data files are shared within the same platform. This avoids
unnecessarily conversions and loss of precision. It is up to the reader
program to do the conversion. I think that is exactly what the "Native format"
is. That one provides speed with the risk that the file is not exchangable
with another platform, or even another MPIO implementation. The canonical
format is supposedly to provide a definite way to do file exchange between
platforms with the potential risk of loss of speed and precision. My feeling
is that couple canonical formats would be sufficient for data exchange purpose.
-- Albert
--